TL;DR
This study introduces a novel radiomics model that leverages the correlation between primary prostate tumor features and metastatic lymph node (MLN) features from biparametric MRI (bpMRI). By incorporating this correlation, the "lesion-correlation model" significantly outperforms a model based solely on primary tumor features in predicting pelvic lymph node metastasis (PLNM), offering a more accurate and potentially less invasive diagnostic tool.
What Problem Does This Paper Address?
Accurately predicting pelvic lymph node metastasis (PLNM) in prostate cancer (PCa) is crucial for determining prognosis and guiding treatment. While MRI has become a valuable tool, conventional MRI methods for detecting PLNM often suffer from low sensitivity. Radiomics, which extracts quantitative features from medical images, has shown promise in improving diagnostic accuracy. However, existing radiomics models for PLNM prediction primarily focus on features from the primary tumor itself. This paper addresses the gap in understanding whether incorporating information from metastatic lymph nodes (MLNs) themselves, specifically their radiomic features and their correlation with primary tumor features, can enhance the predictive power of these models. The hypothesis is that MLNs, being derived from the primary tumor, should share similar metastatic characteristics that can be captured and leveraged for better prediction.
Methods
The researchers developed two radiomics models:
- Lesion Model: This model was built using radiomic features (RFs) extracted solely from primary prostate lesions. Feature selection involved a combination of t-tests and the LASSO (Least Absolute Shrinkage and Selection Operator) algorithm.
- Lesion-Correlation Model: This novel approach first identified primary lesion RFs that were highly correlated (correlation coefficient > 0.9) with MLN RFs. Subsequently, these selected primary lesion RFs underwent further feature selection using t-tests and LASSO to build the model.
Both models were constructed using biparametric MRI (bpMRI) data, specifically T2-weighted imaging (T2WI) and apparent diffusion coefficient (ADC) maps. The models were trained and validated on data from two tertiary hospitals. Support Vector Machine (SVM) classifiers were used for model building, with five-fold cross-validation employed.
Datasets
The study utilized data from 394 prostate cancer patients across two tertiary hospitals. These patients were divided into:
- Training Group: 263 patients (93 PLNM-positive, 170 PLNM-negative).
- Internal Validation Group: 74 patients (19 PLNM-positive, 55 PLNM-negative).
- External Validation Group: 57 patients (14 PLNM-positive, 43 PLNM-negative).
The study did not introduce a new benchmark dataset but rather used retrospective data from these institutions.
Key Insights and Ideas
- Interconnectedness of Primary and Metastatic Lesions: The core idea is that metastatic lymph nodes (MLNs) are not independent entities but rather extensions of the primary tumor. Their radiomic features should therefore reflect the metastatic potential and characteristics of the primary lesion.
- Correlation as a Feature Selection Enhancer: The study highlights that simply extracting features from the primary lesion might miss crucial information. By analyzing the correlation between primary lesion and MLN features, researchers can identify primary lesion features that are more indicative of metastatic potential.
- T2WI Dominance in Feature Importance: Interestingly, both models found that features extracted from T2WI sequences held higher weight coefficients compared to those from ADC maps. This suggests that T2WI might be more effective in capturing the aggressiveness and infiltration characteristics relevant to PLNM prediction.
- Entropy as a Key Predictor: The "T2_tumor_wavelet_LLL_glszm_ZoneEntropy" feature emerged as a highly weighted predictor in the lesion-correlation model, suggesting that textural heterogeneity captured by this feature is strongly associated with PLNM.
Key Results
The lesion-correlation model consistently outperformed the lesion model across all validation groups:
- Training Group:
- Lesion Model AUC: 0.8053
- Lesion-Correlation Model AUC: 0.8466 (p=0.0002)
- Internal Validation Group:
- Lesion Model AUC: 0.7321
- Lesion-Correlation Model AUC: 0.8268 (p=0.0429)
- External Validation Group:
- Lesion Model AUC: 0.6445
- Lesion-Correlation Model AUC: 0.7874 (p=0.0431)
The DeLong test confirmed statistically significant improvements in AUC for the lesion-correlation model in all groups.
Limitations and Future Work
The authors acknowledge several limitations:
- Diagnostic Confirmation of MLNs: In the training group, a minority of PLNM-positive cases were confirmed by biopsy or treatment response, while others were diagnosed by bpMRI. This could introduce some diagnostic variability.
- Exclusion of Deep Features: The study did not extract deep learning-based features from MLNs, which could potentially offer further insights.
- Inconsistent Imaging Devices: The use of imaging devices from different manufacturers (Siemens and Philips) across the two centers might introduce variability in scanner performance and scanning parameters.
Future work could involve:
- Incorporating Deep Learning: Exploring deep learning algorithms for MLN feature extraction.
- Standardizing Imaging Protocols: Conducting studies with standardized imaging protocols to mitigate scanner-related variability.
- Further Validation: Expanding validation to larger, more diverse multi-center datasets.
Summary updated